Decouple ATIF schemas from core and make nvidia-nat-eval lightweight by default

ci/scripts/github/build_wheel.sh

@@ -113,13 +113,17 @@ for whl in "${MOVED_WHEELS[@]}"; do
                 exit ${IMPORT_TEST_RESULT}
             fi
 
-            REPORTED_VERSION=$(nat --version 2>&1)
-            NAT_CMD_EXIT_CODE=$?
-
-            if [[ ${NAT_CMD_EXIT_CODE} -ne 0 ]]; then
-                rapids-logger "Error 'nat --version' command failed exit code ${NAT_CMD_EXIT_CODE} from wheel ${whl} with Python ${pyver}"
-                echo "${REPORTED_VERSION}"
-                exit ${NAT_CMD_EXIT_CODE}
+            if command -v nat >/dev/null 2>&1; then
+                REPORTED_VERSION=$(nat --version 2>&1)
+                NAT_CMD_EXIT_CODE=$?
+
+                if [[ ${NAT_CMD_EXIT_CODE} -ne 0 ]]; then
+                    rapids-logger "Error 'nat --version' command failed exit code ${NAT_CMD_EXIT_CODE} from wheel ${whl} with Python ${pyver}"
+                    echo "${REPORTED_VERSION}"
+                    exit ${NAT_CMD_EXIT_CODE}
+                fi
+            else
+                rapids-logger "Skipping nat CLI test; 'nat' command not installed by wheel ${whl}"
             fi
         else
             rapids-logger "Skipping nat CLI test for nvidia_nat_app (framework-agnostic package); verifying nat_app import"

docs/source/extend/custom-components/custom-evaluator.md

@@ -154,6 +154,9 @@ You can also author a custom evaluator that only implements ATIF-native scoring
 When using `AtifBaseEvaluator`, implement `evaluate_atif_item` and reuse the built-in concurrent `evaluate_atif_fn`.
 This is useful when your scoring logic consumes canonical ATIF trajectories directly.
 
+This example uses evaluator registration (`@register_evaluator`) and therefore requires full runtime dependencies (`nvidia-nat-eval[full]`).
+Base `nvidia-nat-eval` is sufficient for standalone ATIF harness usage without workflow or plugin registration.
+
 The following example registers a minimal ATIF-only cosine-similarity evaluator:
 `examples/evaluation_and_profiling/simple_web_query_eval/src/nat_simple_web_query_eval/atif_only_evaluator_register.py`:
 ```python
@@ -165,8 +168,8 @@ from pydantic import Field
 from nat.builder.builder import EvalBuilder
 from nat.builder.evaluator import EvaluatorInfo
 from nat.cli.register_workflow import register_evaluator
-from nat.data_models.evaluator import EvalOutputItem
 from nat.data_models.evaluator import EvaluatorBaseConfig
+from nat.plugins.eval.data_models.evaluator_io import EvalOutputItem
 from nat.plugins.eval.evaluator.atif_base_evaluator import AtifBaseEvaluator
 from nat.plugins.eval.evaluator.atif_evaluator import AtifEvalSample
 

docs/source/get-started/installation.md

@@ -38,7 +38,8 @@ To install these first-party plugin libraries, you can use the full distribution
 - `nvidia-nat[agno]` or `nvidia-nat-agno` - [Agno](https://agno.com/)
 - `nvidia-nat[crewai]` or `nvidia-nat-crewai` - [CrewAI](https://www.crewai.com/) Conflicts with `nvidia-nat[openpipe-art]`.
 - `nvidia-nat[data-flywheel]` or `nvidia-nat-data-flywheel` - [NeMo DataFlywheel](https://github.com/NVIDIA-AI-Blueprints/data-flywheel)
-- `nvidia-nat[eval]` or `nvidia-nat-eval` - Evaluation orchestration package
+- `nvidia-nat[eval]` or `nvidia-nat-eval[full]` - Full evaluation runtime dependencies for config-driven `nat eval` workflows
+- `nvidia-nat-eval` - Evaluation package for ATIF-native and standalone custom evaluator workflows
 - `nvidia-nat[langchain]` or `nvidia-nat-langchain` - [LangChain](https://www.langchain.com/), [LangGraph](https://www.langchain.com/langgraph)
 - `nvidia-nat[llama-index]` or `nvidia-nat-llama-index` - [LlamaIndex](https://www.llamaindex.ai/)
 - `nvidia-nat[mcp]` or `nvidia-nat-mcp` - [Model Context Protocol (MCP)](https://modelcontextprotocol.io/)

docs/source/improve-workflows/evaluate.md

@@ -23,7 +23,12 @@ NeMo Agent Toolkit provides a set of evaluators to run and evaluate workflows. I
 
 ## Prerequisites
 
-In addition to the base `nvidia-nat` package, you need to install the evaluation package to use `nat eval`. Install the evaluation extra package with one of the following commands, depending on whether you installed the NeMo Agent Toolkit from source or from a package.
+Choose the installation mode that matches your evaluation workflow:
+
+- Standalone ATIF evaluation (`EvaluationHarness` plus ATIF-native custom evaluators): install base `nvidia-nat-eval`.
+- Full `nat eval` runtime (workflow execution, dataset readers such as `csv`/`parquet`/`xls`, and config-driven evaluators): install `nvidia-nat[eval]`.
+
+For source installs:
 
 ::::{tab-set}
 :sync-group: install-tool
@@ -50,6 +55,12 @@ uv pip install "nvidia-nat-eval"
 ::::
 
 
+For package installs, use the NeMo Agent Toolkit `metapackage` to run `nat eval`:
+
+```bash
+uv pip install "nvidia-nat[eval]"
+```
+
 If you plan to run profiling via `nat eval` (for example, when `eval.general.profiler` is enabled), install the profiler package as well:
 
 ::::{tab-set}

docs/source/reference/cli.md

@@ -555,7 +555,11 @@ nat serve --config_file=path/to/config --host 0.0.0.0 --port 8000
 The Swagger API docs will be available at: [http://localhost:8000/docs](http://localhost:8000/docs)
 
 ## Evaluation
-The `nat eval` command is provided by the `nvidia-nat-eval` package. Install evaluation support with `pip install "nvidia-nat[eval]"` or `pip install nvidia-nat-eval`.
+The `nat eval` command is provided by the `nvidia-nat-eval` package.
+
+For full config-driven `nat eval` runtime paths, install `uv pip install "nvidia-nat[eval]"`.
+
+For ATIF-native standalone custom-evaluator paths, install `uv pip install nvidia-nat-eval`.
 
 The `nat eval` command provides access a set of evaluators designed to assessing the accuracy of NeMo Agent Toolkit workflows as
 well as instrumenting their performance characteristics. Please reference

docs/source/resources/migration-guide.md

@@ -58,7 +58,7 @@ This is a breaking change:
 
 To migrate:
 - Install both packages when using these evaluators:
-  - `pip install nvidia-nat-eval nvidia-nat-langchain`
+  - `pip install "nvidia-nat-eval" nvidia-nat-langchain`
 - Install the RAGAS evaluator package when using `_type: ragas`:
   - `pip install nvidia-nat-ragas`
 - Install the profiler package when using performance evaluators or profiling workflows:
@@ -90,7 +90,8 @@ CLI command ownership is now aligned to package domains:
 
 To migrate:
 - Install command-specific packages as needed:
-  - `pip install nvidia-nat-eval`
+  - `pip install "nvidia-nat[eval]"`
+  - `pip install "nvidia-nat-eval[full]"`
   - `pip install nvidia-nat-profiler`
   - `pip install nvidia-nat-security`
 

docs/source/run-workflows/observe/observe-workflow-with-langsmith.md

@@ -94,7 +94,7 @@ The `--override` flag accepts a dot-notation path into the YAML config hierarchy
 ## Structured Evaluation Experiments
 
 :::{note}
-The `nat eval` command is provided by the evaluation package. If the command is not available, install the eval extra first:
+The `nat eval` command is provided by the evaluation package. For full config-driven eval paths, install the full eval extra:
 
 ```bash
 uv pip install -e '.[eval]'